Three Dimensional Microscope Image Super-resolution Using Deep Learning and Self-similarity

Tomoya Komiyama, Yinhao Li, Yen-Wei Chen, Yu Song, Yutaro Iwamoto

Abstract: It has recently become possible to capture three-dimensional volumetric images of plant roots using a microscope. However, due to the limited time available to capture images of living plants’ roots, which are constantly changing, the depth resolution is low. This study aims to create depth-resolution images using super-resolution technology. Existing methods require a pair of high-and low-resolution images to train a model. However, such images are unavailable for the images which need to be developed into higher resolution. In this study, the self-similarity of plant images were used to solve this problem and a convolutional neural network to train on yx-plane images was proposed. The trained model was then used for yz-plane image super-resolution. The proposed method has good performance, according to the experimental results.